Agents as adversaries
How autonomous systems attack, and misbehave, at machine speed — treating the agent itself as the threat, not only its inputs.
AI safety, through security
When agents act,
safety is security.
As AI systems begin to act on the world, the risks show up first as security failures. We study how agents attack, defend the infrastructure they run on, and publish the work.
01 / Agents
02 / Defence
Defending the stack
Detecting and investigating attacks that move faster, and louder, than defenders can follow.
03 / In the open
In the open
Evaluations, detections, data and tools, published so anyone can check the work and build on it.